diffusion-kernel

What problem does it solve?

This Skill provides tools and guidance to optimize the performance of diffusion model kernels, reducing inference latency and improving throughput.

Core Features & Use Cases

• Kernel Development: Guides for writing custom CUDA and Triton kernels for diffusion models.
• Performance Analysis: Tools for profiling and benchmarking kernel performance using Nsight Compute and Triton.
• Optimization Strategies: Best practices for memory access, occupancy, and precision tuning on various GPU architectures (H100, A100, T4).
• Use Case: A developer wants to speed up the RMSNorm operation in a diffusion model. They can follow the add-cuda-kernel.md guide to implement a highly optimized JIT CUDA kernel, test its correctness, benchmark its performance against PyTorch, and profile it with Nsight Compute to ensure it saturates GPU bandwidth.

Quick Start

Use the diffusion-kernel skill to add a new Triton kernel for fused elementwise operations.